Stationarity-Aware Retrieval-Augmented Time Series Forecasting (KDD'26)

This repository provides the official implementation of SARAF, accepted by KDD 2026.

What is SARAF?

Retrieval-augmented forecasting usually assumes that similar historical patterns lead to similar future trajectories. However, this assumption can be unreliable for real-world time series, where different datasets exhibit different levels of stationarity. In highly non-stationary series, such as exchange-rate-like data, two historical windows may look similar in the past but evolve very differently in the future.

SARAF addresses this issue by making retrieval stationarity-aware. Instead of relying only on temporal similarity, SARAF adaptively combines:

• Time-aligned retrieval, which strengthens temporally meaningful historical evidence;
• Diversity-aware retrieval, which avoids redundant neighbors and covers heterogeneous historical regimes;
• Stationarity-aware aggregation, which controls how retrieved futures are fused according to the stationarity of the dataset.

In short, SARAF asks not only:

“Which past segments look similar to the query?”

but also:

“When can their future trajectories be trusted?”

This makes retrieval-augmented forecasting more robust under non-stationary settings while preserving the benefits of similarity-based retrieval on more stable datasets.

---

Required Packages

Install all dependencies:

pip install -r requirements.txt

---

Dataset Preparation

Create a ./data directory and place dataset files inside:

mkdir -p ./data

All standard benchmark datasets (ETT, Electricity, Exchange, Traffic, Solar) can be downloaded from the Autoformer Google Drive.

---

Usage

Run with Scripts (Recommended)

We provide per-dataset bash scripts under ./scripts/. Each script runs experiments across multiple prediction lengths and random seeds.

# ETTh + ETTm (seq_len=720)
bash scripts/ETTh_720.sh

# Electricity
bash scripts/elec_720.sh

# Exchange Rate
bash scripts/exchange_rate_720.sh

# Traffic
bash scripts/traffic_720.sh

# Solar
bash scripts/solar_720.sh

Acknowledgement

This code is based on RAFT and Time-Series-Library. We thank the authors for their open-source contributions.

Citation

If you find this repository useful for your research, please consider citing our paper:

@misc{zhou2026saraf,
  title         = {Stationarity-Aware Retrieval-Augmented Time Series Forecasting},
  author        = {Zhou, Shiqiao and Sch{\"o}ner, Holger and Wu, Zipeng and Fouch{\'e}, Edouard and Wilson, IAG and Wang, Shuo},
  year          = {2026},
  doi           = {10.48550/arXiv.2606.04135},
  url           = {https://arxiv.org/abs/2606.04135}
}